Tencent releases Hy-Nomas: an open video world model for a single RTX 4090

Windows 10 / 11

Windows 10 / 11: not supported. According to the instructions, Hy-Nomas needs a Linux machine with an NVIDIA graphics card (Docker or uv on Linux). There are no instructions for Windows; try the online demo at hy-nomas.hy.tencent.com instead.

macOS

macOS: not supported. The model needs an NVIDIA graphics card with CUDA, which Macs don't have. Use the online demo at hy-nomas.hy.tencent.com.

Linux

Linux: supported. A single NVIDIA GeForce RTX 4090 (24 GB) with the low-mem configuration is the officially tested setup. The manufacturer explicitly recommends the Docker image. The source only shows the prompt file format in the Diffusers example; check the repo if in doubt.

1. Check the prerequisites. You need a Linux machine with an NVIDIA graphics card (target: RTX 4090 with 24 GB), the NVIDIA driver installed, Docker and the NVIDIA Container Toolkit (so Docker can use the GPU). You also need Git, Python 3 and plenty of system RAM, because parts of the model are offloaded to regular RAM. Plan for about 50 GB of free space for the model, the Qwen text model and the Docker image.

   nvidia-smi
   docker --version
   git --version

2. Download the program. Get the code from Tencent and change into the folder.

   git clone https://github.com/Tencent-Hunyuan/Hy-Nomas.git
   cd Hy-Nomas

3. Download only the model files you need. The whole repo is 192 GB. For the 4090 variant, you only need the INT8 DiT and the FP8 decoder (about 9 GB each). They end up in the models folder inside the program folder and keep their subfolders.

   python3 -m venv ~/hf-env
   source ~/hf-env/bin/activate
   pip install -U huggingface_hub
   hf download tencent/Hy-Nomas --local-dir ./models --include "DiT/distilled_int8/*" "VAE/PixelUMM_FT_FP8/*"

4. Pull and start the Docker image. The image contains Python, PyTorch and everything else you need. Run the following commands in the Hy-Nomas folder. Afterwards you're inside the container (a terminal in the container).

   export HY_NOMAS_IMAGE="hynomasoss/hynomasoss:cu130"
   
   docker pull "$HY_NOMAS_IMAGE"
   docker run --rm -it --gpus all --ipc=host --network=host \
     -v "$PWD:/workspace/Hy-Nomas" \
     -w /workspace/Hy-Nomas \
     "$HY_NOMAS_IMAGE" bash

5. Create a sample prompt. Inside the container: save a prompt with the example (a little fox in the forest) as prompts.json. The format matches the example from the instructions.

   cat > prompts.json << 'EOF'
   [
     {
       "Style": "Cinematic fantasy", "perspective_view": "Third-person view",
       "subject_list": [{"subject_id": "fox", "display_name": "a small orange fox", "subject_attribute": "Orange fur, white chest, and a fluffy tail."}],
       "segments": [{
         "visual_description_long": "A small orange fox stands on a stone path in a luminous forest. Blue fireflies drift between the trees.",
         "contained_actions": [{"start_time": "00:00.000", "end_time": "00:05.000", "action_description": "The fox walks slowly along the path."}]
       }]
     }
   ]
   EOF

6. Generate your first video. Start the low-mem configuration. On first run, the program automatically downloads the Qwen text model (Qwen/Qwen2.5-VL-7B-Instruct), which takes a while. Expect long run times, because parts of the model run through system RAM.

   export Hy_Nomas_Model_Path=/workspace/Hy-Nomas/models
   
   bash scripts/inference/inference_from_config.sh configs/experiment/single_gpu_low_mem_inference.yaml \
     dit_weight="${Hy_Nomas_Model_Path}/DiT/distilled_int8/diffusion_pytorch_model_ema_0_995_int8.safetensors" \
     model.vae.checkpoint="${Hy_Nomas_Model_Path}/VAE/PixelUMM_FT_FP8/diffusion_pytorch_model_fp8.safetensors" \
     inference.prompt=/workspace/Hy-Nomas/prompts.json \
     inference.save_path=./outputs/single_gpu

7. Result. The generated video of the little fox in the glowing forest is in the outputs/single_gpu folder.